Probabilistic prediction of vertical deflection for high-speed railway bridges using a Gaussian process

Abstract

Vertical deflection of a high-speed railway bridge is one of the important indicators for managing the safety and running stability of a vehicle. Therefore, efforts have been made to develop sensors for measuring the deflection and predicting its short- and long-term future values. However, the vertical deflection of a railway bridge is stochastic because it involves various sources of uncertainty, which may cause errors in physics-based prediction models. This study proposes a Bayesian approach to build a probabilistic prediction model for the vertical deflection of a railway bridge. For this task, a Gaussian process is introduced to construct a covariance matrix with multiple kernels. Thereafter, actual vision-based measurements, measuring time, and temperature data are used to optimize the hyperparameters of the kernels. As a result, the proposed approach provides a probabilistic prediction interval as well as a predictive mean of the vertical deflections of the bridge. This approach is applied to an actual high-speed railway bridge in the Republic of Korea, and the corresponding analysis results and their performance are discussed.